Why OpenAI and Anthropic are cutting AI model prices by up to 80%

OpenAI and Anthropic have cut prices for several major language models by as much as 80% within months, as competition from each other and from open-weight models puts pressure on the market. Max Woolf writes for Max Woolf’s Blog that the speed and scale of the reductions point to more than a routine product update.

The cuts affect the prices companies pay to process text with AI models, including input tokens, generated output, and cached context. Caching is particularly important for coding agents and other tools that repeatedly reference long conversations or large documents. In those cases, lower cache-read prices can reduce the total cost of a task far more than the headline token price suggests.

OpenAI set the tone in late July when it reduced the price of its GPT-5.6 Luna model by 80%. The model’s input price fell from $1 per million tokens to $0.20, while its output price fell from $6 to $1.20. The company later reduced prices for its larger Sol models as well. By late September, GPT-6.1 Sol cost $2 for input and $10 for output per million tokens, with cache reads priced at $0.10.

Anthropic followed a similar pattern. It made the introductory price for Claude Sonnet 5 permanent, lowered costs for Claude Opus 5.5, and reduced the cache-read price for Fable 5.1 by 75% compared with its predecessor. Anthropic said Opus 5.5 needs less computing power to serve than Opus 5, a claim that offers one possible explanation for the lower pricing.

Competition is changing the economics

According to Woolf, the large AI labs face a market in which users can increasingly switch models with little effort. Coding tools and AI gateways often let customers change providers through a menu option or an API setting. That makes it harder for any single company to maintain a large price premium when rivals offer similar quality.

Open-weight models add further pressure. DeepSeek’s Flash V4.1 model, for example, is available through numerous external inference providers. Those providers run the model on their own computing infrastructure and compete on availability, speed, and price. Woolf notes that 29 providers list the updated model through OpenRouter, up from 13 providers for an earlier version.

This broadens the competitive field beyond OpenAI, Anthropic, Google, and Meta. It also means that a model developer does not fully control the prices customers see. Providers that can run an open model efficiently may undercut one another to win traffic, especially where customers can select the lowest-cost endpoint.

For everyday users, the effect is most visible in AI agents. These tools can generate code, summarize research, draft content, or carry out multi-step tasks. As model prices fall, organizations can run more tasks for the same budget. But lower API prices do not automatically mean lower subscription bills. Woolf points out that OpenAI also reduced usage allowances for one higher-priced Codex plan while introducing a more efficient model.

Lower prices do not settle the profitability question

The timing matters because OpenAI and Anthropic are reportedly approaching potential public offerings. Both companies spend heavily on research, model training, and infrastructure. Price reductions may seem counterintuitive for businesses that need to demonstrate a route to profitability.

Woolf argues that inference, the process of running an already trained model, may still carry healthy margins. If providers can profitably serve tokens at lower prices, competition could continue pushing prices closer to marginal cost. That is the cost of handling one additional unit of usage.

The larger question is whether model quality continues to improve faster than prices decline. Data cited by Woolf from Vercel indicates that average token costs have fallen, while total spending has risen as people use increasingly capable models more often. For content teams and other AI users, the near-term result is clear: switching providers and comparing total task costs, not just list prices, is becoming more important.